Estimating Site Amplification Factor From Microtremor H/v Spectral Ratio by Transfer Learning
Year: 2025
Type: conference-paper
Abstract: Site amplification factors (SAFs) of seismic ground motions are essential in evaluating and predicting seismic hazards. The authors proposed a simple and cost-effective method to estimate a SAF based on a deep neural network (DNN) model and microtremor horizontal-to-vertical spectral ratio (MHVR) observed in Chugoku district, western Japan. This paper aims to develop a transfer learning (TL) model with limited site data for expanding the pre-trained DNN model to the SAFs obtained by different generalized spectral inversion technique (GIT) in different regions. We constructed the TL model that reused most of the neural network layers and their weights from the pre-trained DNN model. The MHVRs and SAFs obtained in eastern regions of Japan were used as the dataset for the TL model. The total data set of 112 sites was divided into training set, validation set, and external test set by 1:1:5. A cross-validation strategy was applied to evaluate the performance of the trained TL model. The degree of fit between the pseudo-SAFs (pSAFs) and the observed SAFs, and the residuals were analyzed for the 80 sites on the test set. The results showed that the accuracy of the TL model with only 16 sites used as the training set outperformed the pre-trained DNN model. The cross-validation results demonstrated that almost consistent prediction results were obtained regardless of any combination of 16 sites selected as the training set. Furthermore, we also compared the impact of different training set sizes on the performance of the TL model. The results indicate that the TL model trained with eight sites showed significant differences in pSAFs among different training set combinations. However, when training with 16 sites, all pSAFs results consistently remained the same. Even though the number of sites was increased to 32, only a slight improvement was observed. We also constructed a DNN model with an extended training set and compared its results with the TL model. The proposed TL model still outperformed the DNN model with the extended training set. Finally, the applicability and limitation of the TL model were discussed by using the MHVRs whose peak frequencies were outside the range of the training set.
Authors: D. Pan, H. Miura
Language: English
Topic: Seismic Waves and Analysis
Subfield: Geophysics
Field: Earth and Planetary Sciences
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7115714081
FWCI:
Cites:
Cited by:
Related to: